fix: append to the vector cache on embed instead of clearing it

[saucam]

Problem (#86)

setEmbedding cleared the entire per-workspace vector cache on every background embed batch. Since batches fire continuously during active work, the next recall() re-read every embedding BLOB from SQLite and re-decoded the full matrix before the cosine loop even started — O(N) I/O + allocation per query that should be O(1), all on the shared daemon event loop.

Fix

Both the id and the vector are already known at write time, so the cached matrix is extended in place instead of being invalidated:

• setEmbedding now takes the workspaceId (the engine has the episode loaded) and upserts the row into the cached matrix.
• Cache entries carry an id → index map, so the re-embed path replaces in O(1) instead of duplicating.
• insert-with-embedding gets the same append treatment (it used to delete the workspace entry).
• If the matrix hasn't been built yet, the write is a no-op on the cache — the next loadVectorMatrix() picks the row up from SQLite with everything else.
• uint8ToFloat32 allocates the Float32Array first and blits bytes into its buffer: one allocation per row instead of two (halves the transient garbage of the remaining cold build).

The cached vector is copied on upsert so later caller mutation can't desync cache from the persisted BLOB.

Complexity

recall() after an embed batch	before	after
SQLite reads	full BLOB table scan (N rows)	0
allocations	2 per row × N	0 (1 small copy per embed write)
per-write cost	O(1) cache clear, deferred O(N) on read	O(1) amortized append

Micro-bench

20,000 episodes × 384-dim embeddings (the issue's ~50k scenario scaled down; scaling is linear), 5 rounds each:

OLD path — full rebuild after embed batch: avg 73.40 ms  (150.4, 50.2, 96.3, 44.6, 25.5)
NEW path — upsert + memoized load:         avg 0.06 ms  (0.19, 0.04, 0.03, 0.02, 0.03)

The old cost recurred on every recall that followed an embed batch — i.e. more or less every recall during active work. It's now paid once per process per workspace (first recall), and never again.

Tests

New describe block pins the behavior via object identity (loadVectorMatrix returns the same object after writes) plus value checks: append on setEmbedding, append on insert-with-embedding, O(1) replace on re-embed without duplication, no-op before first build, and defensive copy semantics. Verified the cached rows match a cold rebuild from a fresh connection.

bun x tsc --noEmit ✓ · bun run lint ✓ · bun run test 756 pass / 0 fail ✓

Related but separate: #71 tracks replacing the brute-force cosine scan itself with an ANN backend.

Fixes #86

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Newly generated embeddings are now normalized before being saved, improving consistency in vector-based features.
  • Vector cache updates are more reliable: newly added or updated embeddings now appear correctly without duplicating entries or falling out of sync.
  • Changes to an embedding input after saving no longer affect stored results.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro Plus

Run ID: fb9d9e1a-47de-4c1b-91ac-cca53cb8843b

📥 Commits

Reviewing files that changed from the base of the PR and between f7487bc and f3dd57a.

📒 Files selected for processing (3)

• src/daemon/memory/engine.ts
• src/daemon/memory/store.ts
• src/tests/memory.test.ts

---

📝 Walkthrough

Walkthrough

This PR normalizes generated embedding vectors before persisting them in the daemon's embedding worker and adds a test suite validating incremental vector-cache behavior in SqliteEpisodeStore, covering insert, setEmbedding, lazy cache initialization, and vector copy immutability.

Changes

Embedding normalization and vector cache tests

Layer / File(s)	Summary
Normalize embeddings before persisting src/daemon/memory/engine.ts	The background embedding worker now normalizes each generated vector before calling #store.setEmbedding, with id, model name, and workspace id passed through unchanged.
Vector cache incrementality tests src/tests/memory.test.ts	New tests verify setEmbedding updates the loaded vector matrix in place, insertions with embeddings append rows, re-embedding replaces rows without duplication, the cache lazily initializes on first loadVectorMatrix, and vectors are copied to prevent external mutation from affecting the cache.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Out of Scope Changes check	⚠️ Warning	The new embedding normalization in engine.ts is unrelated to #86's cache-invalidation fix.	Move the normalization change to a separate PR or explain why it is required for the cache fix.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly states the main change: appending to the vector cache instead of clearing it.
Linked Issues check	✅ Passed	The cache incrementality changes and tests address #86's goal of avoiding full vector-matrix reloads.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch perf/incremental-vector-cache

Warning

There were issues while running some tools. Please review the errors and either fix the tool's configuration or disable the tool if it's a critical failure.

🔧 ESLint

If the error stems from missing dependencies, add them to the package.json file. For unrecoverable errors (e.g., due to private dependencies), disable the tool in the CodeRabbit configuration.

ESLint install failed. For unrecoverable errors, disable the tool in CodeRabbit configuration.

---

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 73.12%. Comparing base (f7487bc) to head (f3dd57a).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #94      +/-   ##
==========================================
+ Coverage   73.06%   73.12%   +0.06%     
==========================================
  Files          65       65              
  Lines       11032    11057      +25     
==========================================
+ Hits         8060     8085      +25     
  Misses       2972     2972              

Flag	Coverage Δ
daemon	73.12% <100.00%> (+0.06%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/daemon/memory/engine.ts	99.52% <100.00%> (+0.01%)	⬆️
src/daemon/memory/store.ts	91.02% <ø> (+0.38%)	⬆️

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.